Abstract
Wireless link scheduling in high-mobility Device-to-Device (D2D) networks is a critical yet challenging task due to the NP-hard nature of sum-rate maximization and the difficulty of obtaining real-time Channel State Information (CSI). This paper introduces a geometric machine learning (G-ML) framework that addresses these challenges by treating interference patterns as points on a Riemannian manifold. By leveraging the underlying geometry of network layouts, the proposed framework achieves near-optimal performance with extreme data efficiency—requiring only ~100 training samples compared to the hundreds of thousands needed by standard deep learning models. Both supervised (G-SVM) and unsupervised (Geometric k-means) algorithms are introduced, demonstrating robust performance in dynamic environments without the need for instantaneous CSI.
Introduction
In modern wireless communication, particularly in autonomous vehicular networks, efficient link scheduling is essential to manage interference and maximize network throughput. Traditional optimization algorithms, such as FPLinQ, rely heavily on accurate, instantaneous CSI. However, in high-mobility scenarios where devices move at highway speeds, CSI changes so rapidly that it is nearly impossible to acquire accurately in real-time.
Recent deep learning (DL) approaches have attempted to map spatial layouts directly to scheduling decisions. While CSI-free, these models are notoriously "data-hungry," often requiring up to 800,000 training samples to generalize effectively. This paper proposes a geometric alternative that utilizes the structural properties of interference to reduce data requirements by several orders of magnitude while maintaining high performance.
Methodology: The Geometric Framework
The core innovation of this work is moving away from standard Euclidean "flat" space and into Riemannian geometry. The framework represents the local interference environment of each D2D pair as a point on a Riemannian manifold.
Manifold Modeling
Instead of feeding raw coordinates into a neural network, the authors model the local graph around each D2D pair based on its $M$ nearest neighbors. This local structure is represented as a Symmetric Positive Definite (SPD) matrix, specifically a regularized Laplacian matrix. These matrices do not live in a flat space; they reside on a curved Riemannian manifold.
Log-Euclidean Metric (LEM)
To measure the "distance" or similarity in interference between D2D pairs on this manifold, the authors use the Log-Euclidean Metric (LEM). This metric is more effective at capturing the non-linear relationships of wireless interference than standard Euclidean distance, allowing the model to "understand" problem constraints much faster than a standard neural network.
Proposed Algorithms
- G-SVM (Supervised): A Geometric Support Vector Machine that uses a specialized kernel based on the LEM to classify links as active ("ON") or inactive ("OFF").
- Geometric k-means (Unsupervised): For scenarios where no labeled training data is available, the authors adapt the $k$-means algorithm to work on the manifold using Fréchet means to cluster links into active and inactive groups.
Experimental Evaluation
The framework was evaluated using simulated D2D network layouts in 500m × 500m areas with the number of pairs ($L$) ranging from 10 to 500. The channel model utilized outdoor ITU-1411 path-loss models, with additional testing on Rayleigh fast-fading scenarios to evaluate robustness.
Benchmarks included:
- FPLinQ: The state-of-the-art optimization algorithm requiring CSI.
- Spatial Deep Learning: A standard DL model requiring 800,000 samples.
- Graph Embedding DNNs: A more efficient DL model requiring 500 samples.
Key Results and Discussion
The G-ML framework demonstrated remarkable performance and efficiency:
- Extreme Data Efficiency: The G-SVM reached 94.84% of the sum rate of FPLinQ using only 90 training layouts. This represents an 82% reduction in training data compared to graph-embedding methods and a several-thousand-fold reduction compared to standard spatial deep learning.
- Robustness in Fast-Fading: In environments where CSI is unreliable, the G-ML methods outperformed traditional greedy algorithms and matched the performance of much more complex deep learning models.
- Scalability: The model maintains high performance (above 90% of optimal) even as the number of D2D pairs increases, with a computational complexity of $O(L^2)$.
- Unsupervised Success: The geometric $k$-means approach proved surprisingly robust, often outperforming supervised methods in high-density scenarios by adapting to the specific interference "shape" of the current transmission period.
Conclusion
By respecting the manifold structure of interference patterns, Geometric Machine Learning provides a powerful shortcut for wireless link scheduling. While the current model focuses on binary decisions and relies on local neighborhood sensitivity, it sets the stage for future work in continuous power control and more complex heterogeneous environments. The significant reduction in data requirements makes this approach highly practical for dynamic, real-world wireless networks.
Summary updated